Generate ideas that aren't the first five obvious ones
Get past the brainstorm everyone already had and into the ideas that actually took some thinking.
- Use it for
- Anyone brainstorming who's tired of getting back the same five generic ideas.
"Give me ideas for X" reliably produces the same handful of ideas anyone would think of in the first thirty seconds, because that's exactly what the prompt asked for and exactly what the training data has the most of — the obvious, most-discussed answers to the most commonly asked version of the question. Those aren't bad ideas, but they're not what you needed a brainstorm for; you could have listed them yourself. The ideas worth paying for are the ones past that first layer, and getting there requires explicitly generating and discarding the obvious ones first, then pushing further.
This prompt forces that discard step instead of letting the first five ideas pass as the final answer.
When not to use this
If the obvious answer is actually the right one — sometimes it is — this is overkill. This is for when you already know the first-thought answers and specifically need something you haven't considered, not for genuine first-pass brainstorming on a brand new problem.
Fill in
I need ideas for: {{prompt_topic}}
Constraints or context: {{constraints}}
Don't give me your first five ideas as the final list. Instead:
1. First, generate the five most obvious ideas — the ones anyone would
think of immediately. List them quickly, don't develop them.
2. Now explicitly set those aside. Generate five more ideas that
approach this from a different angle entirely — a different
audience, a different mechanism, an opposite assumption, a different
scale.
3. From the second batch, pick the two most interesting and develop
them properly — what makes each one actually work, and the biggest
risk or weakness in each.
4. Tell me honestly if any of the "obvious" first five are actually
still the strongest option despite being obvious — obvious doesn't
automatically mean wrong, and I don't want you avoiding a good idea
just to seem original.Why it works
The mechanism is forcing the model past the highest-probability completions, which is where the generic ideas live, into a second pass that's explicitly instructed to diverge. Asking for the obvious ideas first and setting them aside does two things: it clears them out of the way so they don't just get restated with different words, and it gives you a visible baseline to judge the second batch against.
The honesty check at the end matters because the goal isn't "avoid obvious ideas," it's "don't settle for only the obvious ones." An idea being unoriginal doesn't make it wrong, and a prompt that only optimizes for novelty will happily hand you a worse idea because it sounds fresher.
Common mistakes
- ×Skipping straight to asking for "creative" or "unique" ideas without the two-batch structure. That framing alone doesn't reliably produce different output — the explicit discard step is what does the work.
- ×Treating the second batch as automatically better just because it's less obvious. Judge them on merit, using the checkpoint question to keep yourself honest.
- ×Being vague about constraints. Real constraints are what make an idea specific and usable instead of generic in a different way.
Follow-ups
- Combine the strongest elements of two of these ideas into something new.
- What would have to be true for the riskiest idea here to actually work?
Did this work?
At least one idea in the final list is something you hadn't thought of and wouldn't have gotten from a plain "give me ideas for X" prompt.
Tested on claude-opus-5. Evidence status is draft; it moves to battle-tested only on recorded runs, never by hand.